Extended thinking (API)

Extended thinking is a feature of the Claude Messages API that gives Claude a dedicated token budget to reason through complex problems step by step before producing its final response. Instead of jumping straight to an answer, Claude performs multiple sequential reasoning steps internally — a process sometimes called 'serial test-time compute' — and the API returns either a summary or the full record of that reasoning alongside the final text response. The feature has two modes. The older 'manual' mode (type: 'enabled') lets developers set an explicit budget_tokens value to control exactly how many tokens Claude may spend thinking. The newer 'adaptive' mode (type: 'adaptive'), introduced with the Claude 4 family, accepts an effort level (low, medium, high, or max) and lets Claude decide how much reasoning each step actually needs. Adaptive mode is the current standard; manual budget mode is deprecated on Claude Sonnet 4.6 and later and is strictly blocked on Claude Opus 4.7 and 4.8. Extended thinking is particularly effective for mathematics, formal proofs, complex code debugging, multi-step planning, and autonomous agentic workflows. Thinking tokens count toward billing at the same per-token output rate as regular tokens, but the billed count may differ from the token count visible in the response because you are charged for the full thinking generated, not just any summary returned.